Skip to content
Start free
Get App

Why is China giving away its best AI models for free?

Why China releases open-weight AI models, how they differ from closed and true open-source models, and who benefits: businesses needing data privacy and startups seeking lower costs.

Key Takeaways

  • Open-weight models differ fundamentally from closed API models and from true open-source models, and the distinction matters for control and data privacy.
  • The main beneficiaries of Chinese open-weight releases are enterprises with data sensitivity concerns and cost-sensitive startups, not general consumers.
  • Chinese labs can release open weights because their lower cost structures make it a sustainable business choice.
  • Policy debate over open-weight models remains unresolved, with significant disagreement among major US AI and chip industry players.

What the video covers

  • Chinese open-weight model Kimi K3 matched American frontier model performance at a fraction of the cost, sparking headlines about China catching up on open weights.
  • Nvidia CEO Jensen Huang publicly argued open-weight models are the future and core to preserving American AI leadership, while OpenAI and Anthropic warned of security risks from cheap Chinese models.
  • The US administration appears divided on whether to allow, restrict, or ban open-weight models.
  • Closed models are sold as a service behind an API; the provider controls versions, safety, behavior, and pricing, and recoups high training costs.
  • Open-weight models release trained weights, which dictate model behavior, allowing users to run or embed them without permission, but not the training data or code.
  • True open-source models (training data, code, and process fully public) are rare and currently cannot match frontier model quality.
  • Open-weight releases mainly matter to businesses, especially in healthcare, finance, and legal, that want to keep sensitive data inside their own infrastructure.
  • Startups and small tech companies benefit from open-weight models because they cannot afford frontier model pricing and gain better margins.
  • Nvidia promotes open-weight models because it profits from compute demand regardless of which model wins.
  • Chinese labs like DeepSeek have lower cost structures by design, with DeepSeek V3 training reported at about $5.6 million, making open release a viable business strategy.

Chapters

  1. 00:00Chinese open-weight models make headlines
  2. 01:54Why release open models and why Nvidia supports them
  3. 03:07Closed models as a service
  4. 04:25Open-weight vs. true open source explained
  5. 06:03Who benefits: businesses and startups
  6. 07:30Nvidia's stake in the open-weight market
  7. 09:22DeepSeek's cost structure and training economics
  8. 12:14Architecture trade-offs: dense vs. mixture-of-experts

Answers

Questions about this video

What is the difference between open-weight and open-source AI models?

Open-weight models release the trained weights that control model behavior, so users can run or embed them. True open-source models also make the training data, code, and build process public, allowing anyone to rebuild them from scratch. The video notes true open-source models are rare and currently cannot match frontier quality.

Who is most affected by Chinese open-weight model releases?

The video says businesses in healthcare, finance, and legal that need to keep data inside their infrastructure, and startups or small tech companies that cannot afford frontier model pricing, are the groups most affected. Regular consumers who are not developers or tech businesses are unlikely to see changes.

Why would a Chinese AI lab give away its models for free?

According to the video, Chinese labs like DeepSeek have lower underlying costs by design, so releasing open weights fits their business model. DeepSeek V3's training cost was reported at about $5.6 million, and the strategy is not charity.

Full Transcript — Download SRT & Markdown

00:00
Speaker A
For the past 2 weeks, there has been a lot of noise about Chinese open-weight AI models. The Chinese open-weight model Kimi K3 matched the performance of American frontier models at a fraction of the cost for the first time, and that set off a wave of "China's catching up fast on open weights" headlines.
00:18
Speaker A
a wave of China's catching up fast on open weights headlines. The following day, Jensen Huang, the CEO of Nvidia, publishes his first ever personal social media post saying that the openweight models is the future and they should be
00:35
Speaker A
The following day, Jensen Huang, the CEO of Nvidia, published his first-ever personal social media post, saying that open-weight models are the future and should be core to preserving American AI leadership. Then OpenAI and Anthropic sounded alarms in the opposite direction, warning Washington that cheap and powerful Chinese models could create unacceptable security risks, alluding to a possible dystopian future.
00:58
Speaker A
can't make up its mind on whether to allow, to restrict, or to ban openweight models altogether. And the cherry on top of it all is the crash of the ASML stock. Dutch ASML, the key player in chip manufacturing, shared reports that
01:15
Speaker A
The Trump administration is clearly divided and can't make up its mind on whether to allow, restrict, or ban open-weight models altogether. And the cherry on top of it all is the crash of the ASML stock. Dutch ASML, the key player in chip manufacturing, shared reports that a Chinese state-backed firm had begun mass production of lithography machines, which emphasizes once again that China is closing gaps across multiple layers of the AI stack.
01:38
Speaker A
openw weight when open AI and anthropic are pushing back hard? Since when did Chinese AI labs become a charity? And did they? In today's episode of the business of AI, let's dive in.
01:54
Speaker A
But why are they making these models available? Who does this affect? Why is Nvidia a proponent of open weight when OpenAI and Anthropic are pushing back hard? Since when did Chinese AI labs become a charity? And did they? In today's episode of The Business of AI, let's dive in.
02:10
Speaker A
sells them. You come to this bakery and you buy a scone, one or multiple scones.
02:16
Speaker A
First things first, let's get the definition straight. A lot of people use "open weight" and "open source" interchangeably. So what is the difference between open-source, open-weight, and closed models? Imagine a bakery. The bakery makes baked goods and sells them. You come to this bakery and you buy a scone, or multiple scones.
02:31
Speaker A
finished product that is your scone and you can enjoy it in exchange for payment. This is what a closed model is.
02:38
Speaker A
You can't multiply the scone when you bring it home. You can't bake the exact same scone because you don't have the recipe. You can't change anything about the scone. But what you can do is enjoy it. You have access to a finished product, your scone, and you can enjoy it in exchange for payment. This is what a closed model is.
02:52
Speaker A
software as a service model. As a user, you can rent access to this model for a certain price. The reason all Frontier AI models are sold as closed models is because they're staggeringly expensive to make. And when the company that
03:07
Speaker A
Closed models are sold as a service. The company keeps the model fully behind an API. It controls the version. It controls the safety. It can change the behavior. It can change the pricing. A closed model follows the classic software-as-a-service model. As a user, you can rent access to this model for a certain price. The reason all frontier AI models are sold as closed models is that they're staggeringly expensive to make. When the company that produces the model sells it as a finished product, it recoups the costs. So it puts the model behind an API and sells access by the token. No competitor can directly create a copy of the closed model.
03:24
Speaker A
bakery and they say, "Hey, we are starting to offer breakfast bagels. You can have them with salmon, you can have them with eggs, you can have them with ham, or you can buy these bagels and make something that you want at home."
03:36
Speaker A
Now imagine you come to the same bakery and they say, "Hey, we're starting to offer breakfast bagels. You can have them with salmon, you can have them with eggs, you can have them with ham, or you can buy these bagels and make something you want at home." Again, they're not selling the recipe for the bagel. They're selling customization. This is the open-weight model. In AI models, weights control how the model behaves and how it responds to prompts. This is the most critical and the most valuable part of an LLM. It dictates the behavior of any given LLM.
03:54
Speaker A
the most valuable part of an LLM. It dictates the behavior of any given LLM.
03:59
Speaker A
The companies that provide open-weight models, Meta, DeepSeek, Alibaba, Moonshot, and Mistral, give you something you can actually take and download. That something is the trained weights. When you have those weights, you can run them or embed them inside your own product without asking permission.
04:09
Speaker A
That something is the trained weights. When you have those weights, you can run them or embed them inside your own product without asking permission. And now the last one. You come to the bakery and you see a chef standing at the door
04:25
Speaker A
And now the last one. You come to the bakery and see a chef standing at the door handing out printed recipes for every bagel they sell, every variation, every scone, every single product, every customization, every secret you can think of. They're just giving it out for free. This is true open source, and it's extremely rare. There aren't that many examples of real open-source models. The reason is that they are quite different from the traditional understanding of a model, because they're more of a research artifact or a scientific discovery. They are not commercialized at all, because the training data, the training code, and every step of how they were built and trained is public knowledge, and anyone can rebuild these models from scratch.
04:43
Speaker A
many examples of real open- source models. And the reason for that is because they are quite different from a traditional understanding of the model because they're more of a research artifact or a scientific discovery. They are not commercialized at all because
04:58
Speaker A
It is also important to note that as of today, there are no open-source models able to compete with the quality of frontier models. So when you hear all the fuss about Chinese LLMs or Chinese open models, it's important to note that what they are releasing is the weights. It's the open-weight model. It's not true open source. That distinction matters enormously. But why does it matter enormously? Because this distinction affects a very specific group of users, which brings me to the next point. Who does this really affect? Who should care about open weight?
05:07
Speaker A
It is also important to note that as of today there are no open-source models that are able to compete with the quality of Frontier models. So when you hear all the fuss about Chinese LLMs or Chinese open models, it is important to
05:23
Speaker A
First and foremost, this has nothing to do with regular LLM or AI users. If you're not a tech business and you're not a developer, nothing is going to change in your life because of what the Chinese labs are releasing.
05:45
Speaker A
next point. Who does this really affect? Who should care about the open weight? First and foremost, this has nothing to do with the regular LLM or AI users. If you're not a tech business, if you're not a developer, there is nothing that
06:03
Speaker A
The audience that cares about this is businesses. Businesses in healthcare, finance, and legal, the ones that worry day and night about their data leaving their internal infrastructure. With open-weight models, businesses can run the model entirely inside their own walls, which keeps sensitive data local rather than sending it to OpenAI or Anthropic servers. The second group that cares is startups and small tech businesses. Many of them cannot afford the pricing of frontier companies. So think LLM wrappers, niche AI businesses, indie startups, indie apps, basically anyone who builds a real technology business with a real product. If you have a real product and you switch to an open-weight model, you are going to feel the difference in your margins.
06:07
Speaker A
The audience that cares about this is businesses. Businesses in healthcare, in finance, in legal, the ones that worry day and night about their data leaving the infrastructure, their internal infrastructure. With open weight models, businesses can run the model entirely
06:26
Speaker A
The third group is the infrastructure layer: cloud providers and chipmakers. This one may be a surprise. Why them? Because any business that uses an open-weight model still needs infrastructure for that model. They need the cloud, the GPUs, the memory, and all of the wiring to make those model weights work. So the companies that sell infrastructure to closed models will sell it to open-weight models too. The reason Jensen Huang, the CEO of Nvidia of all places, heavily promotes open-weight models is that Nvidia doesn't care who has the best model. It doesn't care about competition in the model market at all. Whether a model is open source, open weight, or closed, every business using these models needs chips to run them. So Nvidia, as well as other chip providers and cloud providers, is the direct beneficiary of the rise of open-weight models.
06:41
Speaker A
afford pricing of the frontier companies. So think LLM rappers, niche AI businesses, indie startups, indie apps, basically anyone who builds a real technology business and has a real product. If you have a real product and you switch to the openw weight model,
06:59
Speaker A
Now, if you're listening and thinking, but why would you give away those weights? What's the point? If DeepSeek gives away its model weights for free, why would anyone pay for DeepSeek at all? If anybody can download the weights, how are they going to make money? Here's how. Imagine a seed, like a tomato seed, for example. Think of the model weights the way you would think of a seed. When DeepSeek gives access to its model weights, it's publishing the instructions for growing this specific crop. Anyone in the world can take those instructions and plant this seed themselves. But owning a seed and owning a harvest are two completely different things. Having the seed doesn't give you the land. It doesn't give you the water, the fertilizer, the greenhouse, or the labor to grow this seed every single day. Same with the AI model. Running an AI model, actually generating output at scale, reliably and cheaply, takes GPU chips, electricity, data centers, engineering staff, enormous te
07:12
Speaker A
Because any business that uses an openweight model, they still need the infrastructure for that model. They need the cloud, the GPUs, the memory, all of the wiring to make this model weights work. So the companies that sell infrastructure to closed models will
07:30
Speaker A
sell it to openweight models too. The reason Jensen Huang, the CEO of Nvidia of all places, heavily promotes the openweight models is because Nvidia doesn't care who has the best model. It doesn't care about the competition of the model market at all. Whether it's
07:49
Speaker A
open source or open weight or closed, every business using these models needs the chips to run. So, Nvidia as well as other chip providers and cloud providers is the direct beneficiary of the rise of openweight models. Now, if you're
08:05
Speaker A
listening to this and thinking, but why would you sell those weights? What's the point? If Deepseek gives away its model weights for free, why would anyone pay for Deep Seek at all? If anybody can download the weights, how are they going
08:22
Speaker A
to make money? Here's how. Imagine a seed, like a tomato seed, for example. Think of the model weights the way you would think of a seed. When Deepseek gives access to its model weights, it's publishing the instructions for growing this specific
08:41
Speaker A
crop. Anyone in the world can take those instructions and plant this seed themselves. But owning a seed and owning a harvest are two completely different things. Having the seed doesn't give you the land. It doesn't give you the water,
09:01
Speaker A
the fertilizer, or the greenhouse or labor to grow this seed every single day. Same with the AI model. Running an AI model, as in actually generating output at scale, reliably and cheaply, takes GPU chips, electricity, data centers, engineering staff, enormous
09:22
Speaker A
technical skill in squeezing maximum output per dollar out of compute. which means that to produce anything meaningful with those seeds, you need to set up a whole farm. Deepseek gives away the seed for free, but having the seed
09:40
Speaker A
doesn't mean you're going to be able to build a farm. And this is exactly the reason why they made those weights available. The weights are the intellectual property, but the way Deepseek makes money is not because they're selling intellectual property.
09:54
Speaker A
They make money on compute at scale. Largecale computing is their product. Giving away the intellectual property costs nothing marginal because the intellectual property became a sunk cost the day the training finished. What it monetizes after the training is finished
10:17
Speaker A
is the industrial capacity to turn that intellectual property into millions of ser requests per second cheaper than everybody else. Every frontier lab selling API access to their model is essentially running a manufacturing business. If you put it bluntly, the
10:35
Speaker A
product is the text output and the unit cost is how much compute it takes to produce that output and the margin is the gap between what they can charge and what it cost them to make it. So the
10:48
Speaker A
economics of this comes down to one thing. Who has the lowest cost to produce the same unit of output? When it comes to American Frontier Labs, think of them as a huge call center where every incoming customer call, no matter
11:06
Speaker A
how simple or complex it is, gets routed to every single employee on the floor.
11:11
Speaker A
Every single agent from an intern to a VP of customer success, has to look at this question, think about it, and only then does one of them actually answer.
11:20
Speaker A
On one hand, it would seem like this would be incredibly wasteful. But that is how dense AI models like OpenAI's family of models or Enthropics, that's how they work. Every query activates the full model, the entire staff, regardless
11:36
Speaker A
of how simple the question is. What Chinese are doing instead is they're adding an intelligent router. The router looks at that incoming question first and sends it directly to only two specialists on the floor of that call center who can answer it while the rest
11:54
Speaker A
of the floor stays idle. And if the rest of the floor stays idle, it costs nothing to have that rest of the floor.
12:02
Speaker A
The architecture that Chinese LLMs use is called mixture of experts. And the math behind it is quite simple. Instead of paying to activate the whole model on every request, it only pays for a small slice of that model relevant to the
12:18
Speaker A
specific question that you're asking. Deepseek's underlying cost per unit is smaller by design, which is why the Chinese LLMs are able to offer openweight models. It makes sense for their business. Training Deep Seek's V3 model ended up being $5.6 million.
12:39
Speaker A
GPT4 over $100 million. The cost difference is only possible because of the architectural decision because if you're using mixture of experts, each query costs less to fulfill. And that is why DeepSk can cost the same customer workload for a fraction of the price
13:00
Speaker A
that GPT would pay. Just look at the numbers. The same workload that runs on Deep Seek's API can cost $80 a month. If the same workload is on GBT4, that would cost about $450. And the exact same volume on clawed opus is $900.
13:19
Speaker A
Again, that is only possible because Deepseek's true cost to produce a million tokens of output sits so far below what OpenAI or Anthropic pay to produce for the same output. But wait, if Deep Seek chose this architecture and
13:35
Speaker A
saved so much money, why are Western labs so much more expensive? And that is because the labs made different choices. The mixture of experts architecture that Chinese openweight models use is essentially a routing system and the routing system
13:55
Speaker A
introduces non-determinism which means that if you ask the same question twice you might be getting slightly different responses. Anthropic claude runs on a different type of architecture also known as dense. For a company like Anthropic, whose entire brand depends on
14:13
Speaker A
safety and consistency, a dense architecture is a safer choice even at a much higher cost. They're making a deliberate trade-off because Enthropic is optimizing for consistency. OpenAI, interestingly, landed on the same logic as Deepseek. They also use a mixture of
14:31
Speaker A
experts model but unlike deepseek they reportedly dedicate some of the workload to safety related tasks trying to get both the cost efficiency and consistency at once. The core thing that is important to understand here is that the architectural choice reflects what each
14:50
Speaker A
lab optimizes for. Deepseek optimized for the cost. It may not be the most accurate but it's free. Enthropic optimized for consistency and that's why it cost so much and OpenAI tried to get both but from many of our previous
15:06
Speaker A
videos you know that OpenAI is really struggling with financials but again this doesn't really answer my question.
15:14
Speaker A
Sure your product can be way cheaper than anthropics but it'll do for the vast majority of LLM rappers. Why offer free weights? And this, ladies and gentlemen, leads me to my final point because there is no real advantage in
15:34
Speaker A
staying closed. Remember the seeds and the farm we talked about? Deep Seek's defensibility never came from the seed or from the model. It came from the farm. In other words, it came from their ability to run the cheapest farm at scale. If a
15:57
Speaker A
competitor cannot undercut your production costs, even if they have the full recipe in hand, if someone steals it, if it gets leaked, that recipe doesn't protect your true value. This is the same reason why Jensen Huang is a
16:12
Speaker A
proponent of open weight. Sure, his company is going to make money off of whatever model they produce, be it openweight or open source or closed. But his main point as to why openw weight is the future, is that even fully closed
16:26
Speaker A
frontier labs cannot actually stop their models from being leaked into the open ecosystem. We made a whole episode on corporate thefts and OpenAI, Antropic, and Gemini and every other Frontier Lab can fall victim to a corporate theft.
16:43
Speaker A
This is very real. So whether the secret gets leaked or if one tries really hard to copy from you or to steal from you, they can query a closed models API repeatedly and use its outputs as training data for their own model. A
17:01
Speaker A
process otherwise known as distillation, which would effectively clone a lot of the core capability without ever seeing the weights. Perhaps the weights should not be your main weapon. Bottom line, if your seeds are spreading anyway, it is time to build an unbeatable farm. So,
17:22
Speaker A
let's come back to where we started. Jensen Huang saying open weights are the future. Open AI and anthropic screaming that they are a security risk because for OpenAI anthropic, their pricing power is their business. So when they warn Washington about dystopian risks
17:41
Speaker A
from cheap Chinese open models, don't fixate just on the word security. Even though security is a very valid concern, all I'm saying is don't forget that behind the security narrative, the big boys are playing big business and big
17:58
Speaker A
geopolitics. We hope this was helpful and we'll see you in the next episode. Bye.
Topics:open-weight modelsopen source AIChinese AIKimi K3DeepSeekNvidiaclosed modelsLLMAI businessAI policy

Get More with the SozAI App

Transcribe recordings, audio files, and YouTube videos — with AI summaries and speaker detection. 30 minutes free.

App Store Google Play
Or transcribe another YouTube video here →

Free tools: TXT to SRT · SRT Validator · Merge SRT · Subtitle to Text · All tools

More from TechButMakeItReal

View All
What happens to Life Insurance and Pensions If AI Bubbl… — Transcript24:32
What happens to Life Insurance and Pensions If AI Bubbl… — Transcript
4,047 1 English
From Sketch to Logo: My Full Process (Real Client Work) — Transcript14:21
From Sketch to Logo: My Full Process (Real Client Work) — Transcript
3,204 1 English
An admonition from Gods Servant Bishop David Oyedepo — Transcript1:39
An admonition from Gods Servant Bishop David Oyedepo — Transcript
191 1 English
Whales Evolved from This Animal — Transcript2:31
Whales Evolved from This Animal — Transcript
555 1 English
CXMT Making Lots of RAM — Transcript0:37
CXMT Making Lots of RAM — Transcript
110 1 English
China’s CXMT Mass-Producing New DRAM—Is Samsung’s Lead Shrinking? — Transcript10:05
China’s CXMT Mass-Producing New DRAM—Is Samsung’s Lead Shrinking? — Transcript
490 1 English
China Just Broke the DDR5 RAM Cartel — $400 RAM Drops to $70 — Transcript9:44
China Just Broke the DDR5 RAM Cartel — $400 RAM Drops to $70 — Transcript
1,558 1 English
When Winter Is Coming, I Always Do This! 14 Freezer Hacks That Save Food & Money — Transcript10:44
When Winter Is Coming, I Always Do This! 14 Freezer Hacks That Save Food & Money — Transcript
2,327 1 English
DO THIS to Purify your Blood Naturally | Dr. Hansaji — Transcript4:00
DO THIS to Purify your Blood Naturally | Dr. Hansaji — Transcript
395 1 English
Water Helps Your Blood Sugar! Dr. Mandell — Transcript0:50
Water Helps Your Blood Sugar! Dr. Mandell — Transcript
77 1 English
Which Gas Station has the Cheesiest Item? — Transcript0:50
Which Gas Station has the Cheesiest Item? — Transcript
197 1 English
How to Naturally Cleanse Your Blood | Dr. Mandell — Transcript13:48
How to Naturally Cleanse Your Blood | Dr. Mandell — Transcript
3,237 1 English